acceptodds
Under review as a conference paper at ICLR 2027

Future-Query Wind Forecasting for Quadrotor Preview Control

Abstract

Quadrotor control can exploit wind preview to anticipate aerodynamic disturbances, yet onboard sensing is sparse and confined to the traversed trajectory, leaving future control queries outside direct sensing support. We introduce EgoCast, a framework for turning sparse onboard observations into short-horizon wind preview at future control queries. Although these queries are unobserved, recent sensing and vehicle motion still provide structure that can be carried toward them. EgoCast therefore decomposes forecasting into a sensing-supported transported reference and a learned residual for unresolved correction and short-horizon evolution; label-anchored physics further constrains this residual where direct supervision is sparse. Under matched training, transport–residual forecasting reduces future-query RMSE by 18.5–23.6% relative to direct full-field prediction. Physical supervision further improves forecasts and benefits 21 of 25 sensing–supervision configurations across shifted observation layouts, with generally larger gains farther from historical observations. These forecast gains reduce aerodynamic force-preview RMSE by 29.1–39.3% relative to the data-only counterpart and improve closed-loop tracking. Adding EgoCast preview to DOB-MPC reduces tracking error on all four tested trajectories relative to DOB-MPC, showing that anticipatory wind preview complements reactive disturbance estimation. Together, these results establish future-query wind forecasting as a practical interface between sparse onboard sensing and preview control.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Wind Forecasting, Sparse Onboard Sensing, Physics-Informed Learning, Preview Control

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_hTqzOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies short-horizon wind prediction at future spatial queries required by quadrotor preview control, where onboard observations are sparse and confined to the vehicle's past trajectory. EgoCast combines a deterministic Ego-Transport reference, constructed from recent observations and extrapolated vehicle motion, with learned zero-delay corrections and future residual evolution. Training further incorporates moving-frame physical constraints, including a label-anchored stencil that inserts existing velocity labels into local continuity and momentum residuals. The predicted wind field is converted into aerodynamic force previews for MPC and is also evaluated in combination with a disturbance observer. Experiments in coupled OpenFOAM–preCICE–MuJoCo simulations compare direct and transport–residual forecasting, physical-supervision variants, sensing/supervision budgets, distance from historical observations, aerodynamic force prediction, and closed-loop tracking.

Critical strengths
  1. The core forecasting design is supported by careful matched comparisons. The direct and transport–residual variants use matched architectures, parameter counts, motion inputs, labelled data, optimizer budgets, and paired seeds. The paper also explicitly acknowledges that this comparison evaluates the complete parameterization, including the deterministic reference and different initial output field, rather than isolating the residual form alone. This makes the comparison informative without overstating what it proves.
  2. The paper performs unusually informative physical-supervision ablations and reports negative results rather than hiding them. The all-predicted-stencil comparison establishes an additional empirical benefit from label insertion beyond ordinary physical regularization. The distance and budget analyses further show that the gain is concentrated in more weakly supported regions and can become small or negative in other regimes. The authors also report that removing the moving-frame terms does not establish an accuracy loss, and the separate reconstruction study shows that physical regularization can oversmooth local structure. These results make the contribution more credible and appropriately bounded.
  3. The evaluation connects forecasting to downstream control instead of stopping at field RMSE. The paper includes aerodynamic force-preview evaluation, four closed-loop trajectories, DOB integration, velocity-noise tests, and computational profiling. The strong Ego-Transport baseline is also retained throughout, which is important because it prevents the learned method from benefiting from an artificially weak control baseline.
Critical weaknesses
  1. The closed-loop experiments do not fully isolate the incremental control value of learned future evolution. In the selected 16/33 setting, Ego-Transport already outperforms EgoCast on the three short trajectories, while the DOB comparison evaluates DOB+EgoCast without a matched DOB+Ego-Transport baseline. Thus, the current results establish the value of the complete wind-informed preview pipeline, but not how much of the gain specifically comes from learned temporal anticipation. In addition, each closed-loop configuration uses only one trained predictor, so downstream training-seed variability is not assessed.
  2. The physical-supervision experiments support an empirical benefit, but the mechanism is not fully isolated. The all-predicted-stencil ablation shows that label insertion provides an additional gain beyond ordinary physical regularization. However, it remains unclear how much of this gain is specific to the proposed PDE-based coupling rather than the changed use of supervised labels. Since the auxiliary pressure and eddy-viscosity fields are unsupervised, basic held-out physical-residual diagnostics would strengthen the interpretation. I view this as a mechanism-attribution limitation rather than evidence that the component is ineffective.
Questions for authors
  1. Control attribution. Can you provide a matched DOB+Ego-Transport comparison under the same controller and trajectories as DOB+EgoCast? This would clarify how much of the downstream gain comes specifically from learned temporal anticipation.
  2. Physical-supervision diagnostics. Can you provide held-out continuity/momentum residuals or basic term-wise diagnostics for the auxiliary pressure and eddy-viscosity contributions? This would help clarify the physical interpretation of the regularizer.
  3. Experimental scope. Please clarify the inflow conditions of the two figure-eight training campaigns in Table 7 and which observation layouts are intended to represent feasible onboard sensing.
Reviewer_wCiNOpenReview ↗
Rating
3(+1)
Confidence
1
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The goal of this paper is to propose a method for estimating future wind conditions in order to anticipate the control required by a moving vehicle. This task is non-trivial, as the observations collected along the vehicle's trajectory differ in both location and time from those needed to predict future wind conditions.

To address this challenge, the proposed method encodes the observation history into a latent context vector ztz_t. This vector is then fed into two neural networks. The first predicts an instantaneous correction to the velocity field, while the second predicts its temporal evolution. Both predictions are treated as corrections to a baseline estimate (called Ego-Transport) obtained by interpolating recent observations. Finally, since the training targets are sparse, the method introduces a spatial consistency loss derived from physical laws and implemented using finite-difference techniques.

Critical strengths
  • The experimental results are promising compared with other learned baselines (DeepONet and Adapted SHRED, see Table 1).
  • Although I am not familiar with this research area, the clear and simple presentation of the problem and the method, together with Figures 1, 2, and 3, helped me understand the main ideas quite easily. However, I believe the presentation could be further improved, particularly in the more technical sections (Sections 3.2 and 3.3). This would make the paper more accessible to a broader community.
Critical weaknesses
  • The scaling behavior of the method with the number of observations MM and supervision targets SS is somewhat counterintuitive. The results given in table 1 are better with M=S=8M=S=8 than with M=S=16M=S=16. However, this may be due to a misunderstanding of the experimental setup on my part.
  • From Tables 1 and 4, Ego-Transport already achieves strong results and appears competitive with the full method (which includes the residual correction terms). Thus, at first glance, the benefit of these residual terms is not obvious to me.
  • All experiments are conducted on simulated data. However, I understand that conducting experiments on real-world systems may be costly and challenging.
Questions for authors
  1. As mentioned above, could the authors clarify the practical benefits of the correction terms (zero-delay and future increment) compared with Ego-Transport alone? In addition, could the authors clarify whether Ego-Transport (i.e. the interpolation algorithm used to construct the reference estimate) is a contribution of this paper or an established method from the existing literature?
  2. Could the authors explain why the methods perform better when M=S=8M=S=8 than when M=S=16M=S=16?
  3. Could the authors discuss potential applications of their method in more detail? The paper frequently refers to a “vehicle” but provides few concrete examples of how the method could be used in practice.

I'd be willing to increase my score if the authors clarify the benefits of the full method over Ego-Transport alone, and particularly whether Ego-Transport itself is a novel contribution or an established method.

Reviewer_LSFHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses short-horizon wind forecasting at unvisited future control queries using sparse onboard quadrotor observations. EgoCast constructs a motion-aware, inverse-distance-weighted Ego-Transport reference and learns zero-delay corrections and gated temporal residuals. Label-anchored physical constraints incorporate sparse velocity labels into local continuity and momentum residuals. Predicted wind is converted into aerodynamic force preview for model predictive control, optionally combined with a disturbance observer.

In coupled OpenFOAM–preCICE–MuJoCo simulations, transport–residual prediction reduces future-query RMSE by 18.5–23.6% relative to matched direct-field predictors. Physical supervision improves 21 of 25 sensing–supervision configurations. Adding EgoCast preview to DOB-MPC also reduces tracking RMSE on all four tested trajectories. The results support the proposed forecasting framework, although closed-loop validation is limited to simulation.

Critical strengths

1. Well-motivated forecasting structure with controlled comparisons.

The distinction between past sensing locations and future control queries is meaningful for onboard wind preview. Equation (2) makes productive use of a transported reference rather than requiring the network to reconstruct the entire future field. Table 1 compares direct and residual prediction under matched architectures, motion inputs, training budgets, and paired seeds. The consistent 18.5–23.6% improvement provides credible evidence for this parameterization.

2. Informative physics ablations and downstream control evaluation.

Tables 2–3 distinguish generic physical regularization from label insertion, showing that the latter particularly helps queries farther from historical observations. Table 4 then connects forecast accuracy to aerodynamic force prediction and tracking, including a disturbance-observer baseline. Importantly, the paper reports cases where Ego-Transport outperforms EgoCast rather than hiding these negative results. This makes the empirical contribution more informative and credible.

Critical weaknesses

1. The incremental control benefit of learned, physics-supervised forecasting is not convincingly isolated.

The simple Ego-Transport reference remains highly competitive. In Table 1, it outperforms EgoCast at M = S = 16 (1.0154 versus 1.0572 m/s). In Table 4(b), Ego-Transport also achieves lower tracking RMSE on all three 5-second trajectories at M = 16, S = 33. At M = S = 8, EgoCast improves over transport, but its tracking errors are nearly identical to those of the data-only residual model across all four trajectories (Table 4(c)). While DOB+EgoCast outperforms DOB-MPC, the paper does not compare it against DOB combined with Ego-Transport. Consequently, the practical advantage attributable to the learned, physics-supervised forecast—rather than simpler spatial transport or residual disturbance compensation—remains unclear. This directly affects the significance of the proposed complexity.

2. Closed-loop validation is narrow, and the reported computation times do not establish real-time feasibility.

Appendix C evaluates closed-loop control under one held-out 300° inflow condition, using a single trained predictor seed per configuration. The four trajectories therefore provide limited evidence about variability across trained models or wind regimes. The additional velocity-noise experiments are useful, but do not address changes in wind dynamics or aerodynamic conditions. Furthermore, Appendix C, Table 21 reports a median EgoCast MPC computation time of 5.63 ms, and 11.44 ms on forecast-refresh updates, whereas the simulated control interval is 4 ms. The controller therefore has not demonstrated the timing required by its stated update schedule on the profiled hardware. These limitations do not invalidate the simulation results, but substantially weaken conclusions about robust, deployable preview control.

Questions for authors

1. Incremental value beyond Ego-Transport.

Could the authors report DOB+Ego-Transport under the same M = 16, S = 33 configuration and four trajectories used for DOB+EgoCast in Table 4? Ideally, DOB+No-PDE would also be included. This would clarify whether the improvement over DOB-MPC requires the learned physics-supervised forecast or can largely be obtained using simpler preview information.

2. Variability of closed-loop improvements.

Appendix C reports closed-loop tracking using only seed 20260826, whereas forecasting uses three paired training seeds. Can the authors provide paired closed-loop tracking results across the existing trained seeds, at least for representative trajectories? This is particularly important at M = S = 8, where the tracking differences between EgoCast and No-PDE are extremely small.

3. Real-time computation budget.

Table 21 reports a median MPC computation time exceeding the 4 ms simulated control interval, even before forecast-refresh overhead. What execution schedule or implementation changes would allow the complete controller to meet its required real-time deadlines? Please distinguish the simulated 250 Hz control updates from experimentally demonstrated wall-clock throughput.

4. Generalization beyond the held-out inflow condition.

The main forecasting test and closed-loop experiments rely on the 300° inflow condition within one CFD environment. Can the authors provide evidence, or existing results, showing how EgoCast performs under a changed wind magnitude, gust frequency, or flow geometry? Such evidence would help determine whether the gains reflect a general forecasting principle rather than specialization to the evaluated flow family.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.